Paper: arXiv 2409.08281
Authors: Shengkun Wang, Taoran Ji, Linhan Wang, Yanshen Sun, Shang-Ching Liu, Amit Kumar, Chang-Tien Lu
Abstract
The stock price prediction task holds a significant role in the financial domain and has been studied for a long time. Recently, large language models (LLMs) have brought new ways to improve these predictions. While recent financial large language models (FinLLMs) have shown considerable progress in financial NLP tasks compared to smaller pre-trained language models (PLMs), challenges persist in stock price forecasting. Firstly, effectively integrating the modalities of time series data and natural language to fully leverage these capabilities remains complex. Secondly, FinLLMs focus more on analysis and interpretability, which can overlook the essential features of time series data. Moreover, due to the abundance of false and redundant information in financial markets, models often produce less accurate predictions when faced with such input data. In this paper, we introduce StockTime, a novel LLM-based architecture designed specifically for stock price data. Unlike recent FinLLMs, StockTime is specifically designed for stock price time series data. It leverages the natural ability of LLMs to predict the next token by treating stock prices as consecutive tokens, extracting textual information such as stock correlations, statistical trends and timestamps directly from these stock prices. StockTime then integrates both textual and time series data into the embedding space. By fusing this multimodal data, StockTime effectively predicts stock prices across arbitrary look-back periods. Our experiments demonstrate that StockTime outperforms recent LLMs, as it gives more accurate predictions while reducing memory usage and runtime costs.
Complexity vs Empirical Score
- Math Complexity: 7.5/10
- Empirical Rigor: 4.0/10
- Quadrant: Lab Rats — high math complexity, limited empirical testing
Why this score: The paper employs advanced deep learning architectures (LLM adaptation, autoregressive encoders, multimodal fusion) with significant mathematical modeling, but the empirical rigor is limited to experimental claims without providing code, datasets, or detailed statistical backtesting metrics.
Research Flowchart
flowchart TD A["Research Goal: Design specialized LLM for stock price prediction"] --> B["Core Methodology: Treat stock prices as consecutive tokens"] B --> C["Data Inputs: Time series data & textual correlations"] C --> D["Multimodal Integration: Fuse data into embedding space"] D --> E["Computational Process: LLM next token prediction"] E --> F["Key Findings: Improved accuracy, reduced memory/runtime"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes StockTime, an LLM-based architecture that treats stock prices as tokens and extracts textual information from them for price prediction.
Source sentence
We introduce StockTime, a novel LLM-based architecture designed specifically for stock price data.
Assumes 5
- The model can rely solely on adjusted close price without any external textual data. data
Source sentence
For Stocktime, we only used the adjusted close price for experiments.
- Freezing the pre‑trained LLM and training only embedding and projection layers is sufficient for good performance. model
Source sentence
and projection layers of the stock time series, we significantly reduce training costs.
- Stock price movements can be modeled as next‑token prediction in an autoregressive fashion. model
Source sentence
It leverages the natural ability of LLMs to predict the next token by treating stock prices as consecutive tokens
- Historical price data from Yahoo Finance for S&P 100 constituents is representative of the market. market
Source sentence
We sourced historical stock price data for S&P 100 companies from Yahoo Finance2 for the period from June 30, 2014, to June 28, 2024.
Show 1 more
- The look‑back window can be selected from {16,32,64,128,256} without loss of generality. model
Source sentence
The lookback window is choosen from {16,32,64,128,256}
Data
- S&P 100 historical prices market-prices public
Source sentence
We sourced historical stock price data for S&P 100 companies from Yahoo Finance2 for the period from June 30, 2014, to June 28, 2024.
- S&P 100 hourly medium‑frequency dataset market-prices public
Source sentence
We also created a hourly medium-frequency stock dataset using companies from the S&P 100, covering the period from June 30, 2023, 9:30 to July 16, 2024, 15:30.
Costs
Not applicable
Finds 2
- StockTime outperforms both general LLMs and FinLLMs on all tested datasets, achieving up to a 5% improvement in stock price movement prediction. performance
Source sentence
Compared to FinLLM, our framework outperformed them on most datasets containing textual data, achieving up to a 5% improvement in stock price movement prediction.
- Ablation shows that removing the autoregressive encoder or the textual information degrades performance, confirming the importance of both components. performance
Source sentence
When the textual information was removed, StockTime’s performance slightly declined, underscoring the importance of integrating multimodal data.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 36 related papers). Coverage is partial, so this is neither replication nor refutation.
4 papers in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (52)
- Blackboard Systems (1986)
- Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education (1983)
- Classification Problem Solving (1984)
- New Ways to Make Microcircuits Smaller (1980) — link
- New Ways to Make Microcircuits Smaller—Duplicate Entry (1980)
- Strategic explanations for a diagnostic consultation system (1984) — link
- Strategic Explanations in Consultation—Duplicate (1983)
- Poligon: A System for Parallel Problem Solving (1986)
and 44 more
- Transfer of Rule-Based Expertise through a Tutorial Dialogue (1979)
- The Engineering of Qualitative Models (2021)
- Attention Is All You Need (2017) — link
- Pluto: The ‘Other’ Red Planet (2015)
- Bloomberggpt: A large language model for finance (2023) — link
- FinBERT: Financial Sentiment Analysis with Pre-trained Language Models (2019) — link
- When flue meets flang: Benchmarks and large pre-trained language model for financial domain (2022) — link
- Bert: Pre-training of deep bidirectional transformers for language understanding (2018) — link
- Electra: Pre-training text encoders as discriminators rather than generators (2020) — link
- Fingpt: Open-source financial large language models — → 2306.06031
- Pixiu: A large language model, instruction data and evaluation benchmark for finance (2023) — link
- Ploutos: Towards interpretable stock movement prediction with financial large language model — → 2403.00782
- A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges — → 2406.11903
- Large language models: A survey (2024) — link
- Autotimes: Autoregressive time series forecasters via large language models (2024) — link
- Time-LLM : Time series forecasting by reprogramming large language models (2024)
- Pytorch: An imperative style, high-performance deep learning library (2019)
- Adam: A method for stochastic optimization (2015)
- Stock price pattern recognition-a recurrent neural network approach (1990)
- A survey of large language models in finance (finllms) (2024) — link
- Stock Movement and Volatility Prediction from Tweets, Macroeconomic Factors and Historical Prices (2023)
- A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (2023)
- Accurate stock movement prediction with self-supervised learning from sparse noisy tweets (2022)
- Stock movement prediction from tweets and historical prices (2018)
- Hybrid deep sequential modeling for social text-driven stock prediction (2018)
- itransformer: Inverted transformers are effective for time series forecasting (2024)
- Alpha-gpt: Human-ai interactive alpha mining for quantitative investment — → 2308.00016
- Enhancing Stock Movement Prediction with Adversarial Training. (2019)
- Add: augmented disentanglement distillation framework for improving stock trend forecasting (2020) — link
- MASTER: Market-Guided Stock Transformer for Stock Price Forecasting (2024)
- Deep learning-based stock price prediction using LSTM and bi-directional LSTM model (2020)
- Combined deep learning classifiers for stock market prediction: integrating stock price and news sentiments (2023)
- ALERTA-Net: A Temporal Distance-Aware Recurrent Networks for Stock Movement and Volatility Prediction (2023)
- Shortlisting machine learning-based stock trading recommendations using candlestick pattern recognition (2023)
- AMA-LSTM: Pioneering Robust and Fair Financial Audio Analysis for Stock Volatility Prediction — → 2407.18324
- Large language models are zero-shot time series forecasters (2024)
- Lag-llama: Towards foundation models for time series forecasting (2023) — link
- Promptcast: A new prompt-based learning paradigm for time series forecasting (2023)
- The wall street neophyte: A zero-shot analysis of chatgpt over multimodal stock movement prediction challenges — → 2304.05351
- One fits all: Power general time series analysis by pretrained lm (2023)
- Mistral 7B (2023) — link
- The Llama 3 Herd of Models (2024) — link
- StockMixer: A Simple Yet Strong MLP-Based Architecture for Stock Price Forecasting (2024)
- Large language models for time series: A survey (2024) — link
6 of 52 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.